End of Day
End of Day
Humans stop working when they get tired. Agents stop working when their context window fills up. Both are hitting a cognitive limit - the inability to hold more information while maintaining quality.
The Context Window as a Workday
A 200K token context window is roughly equivalent to a full workday of agent activity. The agent starts fresh, accumulates observations, makes decisions, takes actions, and records results. Each operation adds tokens. Eventually the window fills and the agent's performance degrades - it starts forgetting early context, making contradictory decisions, repeating work.
This is not a hardware limitation. It is an architectural one. The agent's working memory has a fixed size. When it is full, the agent needs to "sleep" - save its state, clear the context, and start a new session with only the essential carry-forward information.
Designing for the End
Good agent architecture anticipates the end of the context window. It prioritizes information that will be needed later. It summarizes completed work before it scrolls out of context. It saves state to persistent storage before the window fills, not after quality has degraded.
The worst pattern is an agent that runs until it hits the limit and then crashes. The best pattern is an agent that monitors its own context usage, initiates a graceful handoff to a fresh session, and resumes seamlessly.
What Carries Over
The handoff between sessions is where agent quality lives. What gets summarized and carried forward defines the agent's continuity. Too little carry-forward and the agent forgets important context. Too much and the new session starts already half-full.
The art is in the compression: distill a full day of agent work into a few hundred tokens of context that let the next session start as if it remembered everything.
Fazm is an open source macOS AI agent. Open source on GitHub.